The method
Diagnose before prescribing. Every time, including when it is inconvenient.
Clarify the outcome. Simplify. Standardize. Automate deterministically. Add AI only where judgment or language warrants it. Keep human authority explicit. Measure.
How an engagement runs
Five phases, about three weeks, one workflow.
- 01
Phase 0. Fit and authority
One workflow, one executive sponsor, one measurable consequence, and access to the people who actually do the work. If any of those four is missing we say so and stop, because everything after this depends on them.
- 02
Phase 1. Standardized evidence
Role specific intake from the sponsor, the workflow owner, and the operators. Normal cases and exception cases, collected the same way every time, with provenance attached to each piece so a conclusion can be traced back to what it rests on.
- 03
Phase 2. Human led analysis
The evidence is packaged deterministically, then read by people. Every material conclusion cites an evidence identifier. No model decides what your problem is.
- 04
Phase 3. Responsibility design
Step by step: human, deterministic software, AI assisted, or an external system that already owns the answer. Then the controls, approvals, and evidence each automated step requires before it acts.
- 05
Phase 4. A decision, not a pitch
A ranked recommendation with its economics, its risks, its assumptions, and explicit stop conditions. Implementation is a separate decision you make afterwards, with the numbers in front of you.
The six way classification
AI is one of six answers, and usually not the first one.
Every step in your workflow gets sorted into one of these, with the reasoning written down so you can argue with it.
| Step | Owner | Why |
|---|---|---|
| Eliminate | Human | The cheapest automation is the step that stops happening. This is checked first, every time. |
| Standardize | Human | A process nobody agreed on cannot be automated, only accelerated into the same disagreement. |
| Deterministic software | Deterministic | Rules, routing, calculation, validation. Testable, cheap, and it fails the same way twice. |
| AI assisted | AI-assisted | Language and judgment on unstructured input. The model proposes, a person disposes. |
| External system | External system | The system that owns the record answers the question. Copying that state creates a second truth. |
| Leave it human | Human | Consequential, contested, or relationship bearing. Some work should stay with a person who can be asked why. |
What you get
A decision package you own.
- The target workflow and the business outcome attached to it
- A current state map: trigger to completion, including what actually happens
- The systems and data involved, and which one is authoritative when they disagree
- Roles, handoffs, exceptions, and the points where things fail
- Step by step classification across the six categories above
- Risk, evidence, review, and approval requirements for anything automated
- The economics, with assumptions stated and labeled
- A sequenced roadmap with explicit stop conditions
Implementation is excluded, deliberately. Separating the decision from the build is what protects the quality of the recommendation, and it is why we can tell you the answer is not AI without it costing us the project.
Start with one workflow.
Thirty minutes, no deck. Bring a process that costs more than it should and we will tell you honestly whether there is something here worth paying for.